Vehicle Detection¶
The goals / steps of this project are the following:
- Perform a Histogram of Oriented Gradients (HOG) feature extraction on a labeled training set of images and train a classifier Linear SVM classifier
- Optionally, you can also apply a color transform and append binned color features, as well as histograms of color, to your HOG feature vector.
- Note: for those first two steps don't forget to normalize your features and randomize a selection for training and testing.
- Implement a sliding-window technique and use your trained classifier to search for vehicles in images.
- Run your pipeline on a video stream (start with the test_video.mp4 and later implement on full project_video.mp4) and create a heat map of recurring detections frame by frame to reject outliers and follow detected vehicles.
- Estimate a bounding box for vehicles detected.
In [1]:
# Importing some useful packages
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import numpy as np
import cv2
%matplotlib inline
In [2]:
# Inspect the training samples
import glob
vehicle_images = glob.glob("training_images/vehicles/**/*.png")
nb_vehicle_images = len(vehicle_images)
non_vehicle_images = glob.glob("training_images/non-vehicles/**/*.png")
nb_non_vehicle_images = len(non_vehicle_images)
rows = 6
columns = 6
plt.figure(figsize=(20, 9))
for i, file in enumerate(np.random.choice(vehicle_images, 18)):
plt.subplot(rows, columns, i + 1)
plt.axis("off")
plt.title("vehicle")
plt.imshow(cv2.cvtColor(cv2.imread(file), cv2.COLOR_BGR2RGB))
for i, file in enumerate(np.random.choice(non_vehicle_images, 18)):
plt.subplot(rows, columns, i + 1 + 18)
plt.axis("off")
plt.title("non vehicle")
plt.imshow(cv2.cvtColor(cv2.imread(file), cv2.COLOR_BGR2RGB))
In [95]:
# Return the hog features with and optionaly a visualization
# This function is copied from the section 20 of the course
from skimage.feature import hog
def get_hog_features(img, orient, pix_per_cell, cell_per_block, vis=False, feature_vec=True):
if vis == True:
features, hog_image = hog(img, orientations=orient, pixels_per_cell=(pix_per_cell, pix_per_cell),
cells_per_block=(cell_per_block, cell_per_block), transform_sqrt=False,
visualise=True, feature_vector=False, block_norm='L2-Hys')
return features, hog_image
else:
features = hog(img, orientations=orient, pixels_per_cell=(pix_per_cell, pix_per_cell),
cells_per_block=(cell_per_block, cell_per_block), transform_sqrt=False,
visualise=False, feature_vector=feature_vec, block_norm='L2-Hys')
return features
In [96]:
# Test HOG features
test_vehicle = cv2.imread(vehicle_images[0])
test_vehicle = cv2.cvtColor(test_vehicle, cv2.COLOR_RGB2GRAY)
test_non_vehicle = cv2.imread(non_vehicle_images[0])
test_non_vehicle = cv2.cvtColor(test_non_vehicle, cv2.COLOR_RGB2GRAY)
orient = 9
pix_per_cell = 8
cell_per_block = 2
_, vehicle_hog = get_hog_features(test_vehicle, orient, pix_per_cell, cell_per_block, vis = True)
_, non_vehicle_hog = get_hog_features(test_non_vehicle, orient, pix_per_cell, cell_per_block, vis = True)
plt.figure(figsize=(9, 9))
plt.subplot(2, 2, 1)
plt.imshow(test_vehicle, cmap="gray")
plt.axis("off")
plt.subplot(2, 2, 2)
plt.imshow(vehicle_hog, cmap="gray")
plt.axis("off")
plt.subplot(2, 2, 3)
plt.imshow(test_non_vehicle, cmap="gray")
plt.axis("off")
plt.subplot(2, 2, 4)
plt.imshow(non_vehicle_hog, cmap="gray")
plt.axis("off")
Out[96]:
In [5]:
### Training the classifier
In [6]:
# Define a function to extract features from a list of images
# Have this function call bin_spatial() and color_hist()
# This function is copied from section 29 of the class
def extract_features(imgs, cspace='RGB', orient=9,
pix_per_cell=8, cell_per_block=2, hog_channel=0):
# Create a list to append feature vectors to
features = []
# Iterate through the list of images
for file in imgs:
# Read in each one by one
image = cv2.imread(file)
# apply color conversion if other than 'RGB'
if cspace != 'RGB':
if cspace == 'HSV':
feature_image = cv2.cvtColor(image, cv2.COLOR_RGB2HSV)
elif cspace == 'LUV':
feature_image = cv2.cvtColor(image, cv2.COLOR_RGB2LUV)
elif cspace == 'HLS':
feature_image = cv2.cvtColor(image, cv2.COLOR_RGB2HLS)
elif cspace == 'YUV':
feature_image = cv2.cvtColor(image, cv2.COLOR_RGB2YUV)
elif cspace == 'YCrCb':
feature_image = cv2.cvtColor(image, cv2.COLOR_RGB2YCrCb)
else: feature_image = np.copy(image)
# Call get_hog_features() with vis=False, feature_vec=True
if hog_channel == 'ALL':
hog_features = []
for channel in range(feature_image.shape[2]):
hog_features.append(get_hog_features(feature_image[:,:,channel],
orient, pix_per_cell, cell_per_block,
vis=False, feature_vec=True))
hog_features = np.ravel(hog_features)
else:
hog_features = get_hog_features(feature_image[:,:,hog_channel], orient,
pix_per_cell, cell_per_block, vis=False, feature_vec=True)
# Append the new feature vector to the features list
features.append(hog_features)
# Return list of feature vectors
return features
In [7]:
from sklearn.model_selection import train_test_split
from sklearn.svm import LinearSVC
from sklearn.preprocessing import StandardScaler
# colorspace can be one of Can be RGB, HSV, LUV, HLS, YUV, YCrCb
# hog_channel can be one of 0, 1, 2, or "ALL"
# Adapted from the section 29 of the class
def train_svc(colorspace, hog_channel, orient = 9, pix_per_cell = 8, cell_per_block = 2):
sample_size = 4000
cars = vehicle_images[0:sample_size]
notcars = non_vehicle_images[0:sample_size]
car_features = extract_features(cars, cspace=colorspace, orient=orient,
pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
hog_channel=hog_channel)
notcar_features = extract_features(notcars, cspace=colorspace, orient=orient,
pix_per_cell=pix_per_cell, cell_per_block=cell_per_block,
hog_channel=hog_channel)
# Create an array stack of feature vectors
X = np.vstack((car_features, notcar_features)).astype(np.float64)
# Fit a per-column scaler
X_scaler = StandardScaler().fit(X)
# Apply the scaler to X
scaled_X = X_scaler.transform(X)
# Define the labels vector
y = np.hstack((np.ones(len(car_features)), np.zeros(len(notcar_features))))
# Split up data into randomized training and test sets
rand_state = np.random.randint(0, 100)
X_train, X_test, y_train, y_test = train_test_split(
scaled_X, y, test_size=0.2, random_state=rand_state)
# Use a linear SVC
svc = LinearSVC()
# Check the training time for the SVC
svc.fit(X_train, y_train)
# Return the accuracy on the test partition
return round(svc.score(X_test, y_test), 6), svc, X_scaler
In [ ]:
# Checking all the combination of color spaces / channels all other params being fixed
# The idea being to select those 2 params before tuning others
# NOTE: using a single channel is best for performance
for color_space in ('RGB', 'HSV', 'LUV', 'HLS', 'YUV', 'YCrCb'):
for channel in (0, 1, 2, 'ALL'):
score, svc, X_scaler = train_svc(color_space, channel)
print('Score for channel {} of {} = {}'.format(channel, color_space, score))
'''
Results:
Score for channel 0 of RGB = 0.95
Score for channel 1 of RGB = 0.93375
Score for channel 2 of RGB = 0.93375
Score for channel ALL of RGB = 0.960625
Score for channel 0 of HSV = 0.959375
Score for channel 1 of HSV = 0.9075
Score for channel 2 of HSV = 0.938125
Score for channel ALL of HSV = 0.99625
Score for channel 0 of LUV = 0.92375
Score for channel 1 of LUV = 0.90125
Score for channel 2 of LUV = 0.939375
Score for channel ALL of LUV = 0.9925
Score for channel 0 of HLS = 0.956875
Score for channel 1 of HLS = 0.95
Score for channel 2 of HLS = 0.89375
Score for channel ALL of HLS = 0.99625
Score for channel 0 of YUV = 0.943125
Score for channel 1 of YUV = 0.88875
Score for channel 2 of YUV = 0.946875
Score for channel ALL of YUV = 0.999375
Score for channel 0 of YCrCb = 0.93625
Score for channel 1 of YCrCb = 0.87125
Score for channel 2 of YCrCb = 0.92625
Score for channel ALL of YCrCb = 0.990625
'''
In [ ]:
# We will use the H channel of HSL: Score for channel 0 of HLS = 0.956875
# Check if tuning the other parameters is helpful
for orient in (8, 9, 10):
for pix_per_cell in (6, 8, 10):
for cell_per_block in (1, 2):
score, svc, X_scaler = train_svc('HLS', 0)
print('orient {} pix_per_cell {} cell_per_block {} -> {}'.format(orient, pix_per_cell, cell_per_block, score))
# Best result
# orient 9 pix_per_cell 10 cell_per_block 2 -> 0.9575
'''
Results:
orient 8 pix_per_cell 6 cell_per_block 1 -> 0.9575
orient 8 pix_per_cell 6 cell_per_block 2 -> 0.954375
orient 8 pix_per_cell 8 cell_per_block 1 -> 0.95875
orient 8 pix_per_cell 8 cell_per_block 2 -> 0.9525
orient 8 pix_per_cell 10 cell_per_block 1 -> 0.953125
orient 8 pix_per_cell 10 cell_per_block 2 -> 0.9525
orient 9 pix_per_cell 6 cell_per_block 1 -> 0.955625
orient 9 pix_per_cell 6 cell_per_block 2 -> 0.946875
orient 9 pix_per_cell 8 cell_per_block 1 -> 0.954375
orient 9 pix_per_cell 8 cell_per_block 2 -> 0.9525
orient 9 pix_per_cell 10 cell_per_block 1 -> 0.9525
orient 9 pix_per_cell 10 cell_per_block 2 -> 0.9575
orient 10 pix_per_cell 6 cell_per_block 1 -> 0.96
orient 10 pix_per_cell 6 cell_per_block 2 -> 0.94875
orient 10 pix_per_cell 8 cell_per_block 1 -> 0.941875
orient 10 pix_per_cell 8 cell_per_block 2 -> 0.9575
orient 10 pix_per_cell 10 cell_per_block 1 -> 0.9575
orient 10 pix_per_cell 10 cell_per_block 2 -> 0.950625
'''
In [238]:
def draw_boxes(img, bboxes, color=(0, 0, 255), thick=2):
# Make a copy of the image
imcopy = np.copy(img)
# Iterate through the bounding boxes
for bbox in bboxes:
# Draw a rectangle given bbox coordinates
cv2.rectangle(imcopy, bbox[0], bbox[1], color, thick)
# Return the image copy with boxes drawn
return imcopy
# Define a single function that can extract features using hog sub-sampling and make predictions
# This function is adapted from the section 35 of the course to used the previsouly tuned HOG params
def find_cars(img, ystart, ystop, scale, cspace, hog_channel, svc, X_scaler, orient,
pix_per_cell, cell_per_block):
rectangles = []
rectangles_found = []
#img = img.astype(np.float32)/255
img_tosearch = img[ystart:ystop,:,:]
ctrans_tosearch = np.copy(img_tosearch)
# Convert color space
if cspace != 'RGB':
if cspace == 'HSV':
ctrans_tosearch = cv2.cvtColor(img_tosearch, cv2.COLOR_RGB2HSV)
elif cspace == 'LUV':
ctrans_tosearch = cv2.cvtColor(img_tosearch, cv2.COLOR_RGB2LUV)
elif cspace == 'HLS':
ctrans_tosearch = cv2.cvtColor(img_tosearch, cv2.COLOR_RGB2HLS)
elif cspace == 'YUV':
ctrans_tosearch = cv2.cvtColor(img_tosearch, cv2.COLOR_RGB2YUV)
elif cspace == 'YCrCb':
ctrans_tosearch = cv2.cvtColor(img_tosearch, cv2.COLOR_RGB2YCrCb)
else:
ctrans_tosearch = np.copy(img_tosearch)
# rescale image if other than 1.0 scale
if scale != 1:
imshape = ctrans_tosearch.shape
ctrans_tosearch = cv2.resize(ctrans_tosearch, (np.int(imshape[1]/scale), np.int(imshape[0]/scale)))
# Select HOG channel
if hog_channel == 'ALL':
ch1 = ctrans_tosearch[:,:,0]
ch2 = ctrans_tosearch[:,:,1]
ch3 = ctrans_tosearch[:,:,2]
else:
ch1 = ctrans_tosearch[:,:,hog_channel]
# Define blocks and steps as above
nxblocks = (ch1.shape[1] // pix_per_cell) - cell_per_block + 1
nyblocks = (ch1.shape[0] // pix_per_cell) - cell_per_block + 1
nfeat_per_block = orient*cell_per_block**2
# 64 was the orginal sampling rate, with 8 cells and 8 pix per cell
window = 64
nblocks_per_window = (window // pix_per_cell) - cell_per_block + 1
cells_per_step = 2 # Instead of overlap, define how many cells to step
nxsteps = (nxblocks - nblocks_per_window) // cells_per_step
nysteps = (nyblocks - nblocks_per_window) // cells_per_step
# Compute individual channel HOG features for the entire image
hog1 = get_hog_features(ch1, orient, pix_per_cell, cell_per_block, feature_vec=False)
if hog_channel == 'ALL':
hog2 = get_hog_features(ch2, orient, pix_per_cell, cell_per_block, feature_vec=False)
hog3 = get_hog_features(ch3, orient, pix_per_cell, cell_per_block, feature_vec=False)
for xb in range(nxsteps):
for yb in range(nysteps):
ypos = yb*cells_per_step
xpos = xb*cells_per_step
# Extract HOG for this patch
hog_feat1 = hog1[ypos:ypos+nblocks_per_window, xpos:xpos+nblocks_per_window].ravel()
if hog_channel == 'ALL':
hog_feat2 = hog2[ypos:ypos+nblocks_per_window, xpos:xpos+nblocks_per_window].ravel()
hog_feat3 = hog3[ypos:ypos+nblocks_per_window, xpos:xpos+nblocks_per_window].ravel()
hog_features = np.hstack((hog_feat1, hog_feat2, hog_feat3))
else:
hog_features = hog_feat1
xleft = xpos*pix_per_cell
ytop = ypos*pix_per_cell
test_features = X_scaler.transform(hog_features.reshape(1, -1))
test_prediction = svc.predict(test_features.reshape(1, -1))
xbox_left = np.int(xleft*scale)
ytop_draw = np.int(ytop*scale)
win_draw = np.int(window*scale)
rect = ((xbox_left, ytop_draw+ystart),(xbox_left+win_draw,ytop_draw+win_draw+ystart))
rectangles.append(rect)
if test_prediction == 1:
rectangles_found.append(rect)
return rectangles, rectangles_found
In [239]:
# Quick check on still images
# -> using all the channels from the HLS space gives the best results
color_space = 'HLS'
hog_channel = 'ALL'
orient = 9
pix_per_cell = 10
cell_per_block = 2
score, svc, X_scaler = train_svc(color_space, hog_channel, orient, pix_per_cell, cell_per_block)
In [240]:
ystart = 400
ystop = 656
scale = 1.5
rows = 6
columns = 1
plt.figure(figsize=(40, 45))
for i in [1, 2, 3, 4, 5, 6]:
img = cv2.imread('test_images/test{}.jpg'.format(i))
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
plt.subplot(rows, columns, i)
plt.axis("off")
img = draw_boxes(img, rectangles, (120, 120, 120), 1)
img = draw_boxes(img, rectangles_found)
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
In [241]:
# Tune the parameters for the bounding box size
# The larger the car should be the larger the bounding box
img = cv2.imread('test_images/test1.jpg')
out_img = img.copy()
bboxes = []
# Scale 1
ystart = 400
ystop = 600
scale = 1
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
out_img = draw_boxes(out_img, rectangles, (250, 150, 150), 1)
out_img = draw_boxes(out_img, rectangles_found, (250, 150, 150), 6)
bboxes = bboxes + rectangles_found
# Scale 1.5
ystart = 440
ystop = 650
scale = 1.5
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
out_img = draw_boxes(out_img, rectangles, (150, 250, 150), 1)
out_img = draw_boxes(out_img, rectangles_found, (150, 250, 150), 6)
bboxes = bboxes + rectangles_found
# Scale 2
ystart = 500
ystop = 720
scale = 2
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
out_img = draw_boxes(out_img, rectangles, (150, 150, 250), 1)
out_img = draw_boxes(out_img, rectangles_found, (150, 150, 250), 6)
bboxes = bboxes + rectangles_found
plt.figure(figsize=(40, 45))
plt.axis("off")
plt.imshow(cv2.cvtColor(out_img, cv2.COLOR_BGR2RGB))
Out[241]:
In [247]:
# Reduce/elimate false positive
# Use the heatmap solution described in the course
from scipy.ndimage.measurements import label
# Read in image similar to one shown above
img = cv2.imread('test_images/test1.jpg')
heat = np.zeros_like(img[:,:,0]).astype(np.float)
def add_heat(heatmap, bbox_list):
# Iterate through list of bboxes
for box in bbox_list:
# Add += 1 for all pixels inside each bbox
# Assuming each "box" takes the form ((x1, y1), (x2, y2))
heatmap[box[0][1]:box[1][1], box[0][0]:box[1][0]] += 1
# Return updated heatmap
return heatmap
def apply_threshold(heatmap, threshold):
# Zero out pixels below the threshold
heatmap[heatmap <= threshold] = 0
# Return thresholded map
return heatmap
def draw_labeled_bboxes(img, labels):
# Iterate through all detected cars
for car_number in range(1, labels[1]+1):
# Find pixels with each car_number label value
nonzero = (labels[0] == car_number).nonzero()
# Identify x and y values of those pixels
nonzeroy = np.array(nonzero[0])
nonzerox = np.array(nonzero[1])
# Define a bounding box based on min/max x and y
bbox = ((np.min(nonzerox), np.min(nonzeroy)), (np.max(nonzerox), np.max(nonzeroy)))
# Draw the box on the image
cv2.rectangle(img, bbox[0], bbox[1], (0,0,255), 6)
# Return the image
return img
# Add heat to each box in box list
heat = add_heat(heat, bboxes)
# Apply threshold to help remove false positives
heat = apply_threshold(heat,3)
# Visualize the heatmap when displaying
heatmap = np.clip(heat, 0, 255)
# Find final boxes from heatmap using label function
labels = label(heatmap)
draw_img = draw_labeled_bboxes(np.copy(img), labels)
plt.figure(figsize=(40, 45))
plt.subplot(121)
plt.imshow(cv2.cvtColor(draw_img, cv2.COLOR_BGR2RGB))
plt.title('Car Positions')
plt.subplot(122)
plt.imshow(heatmap, cmap='hot')
plt.title('Heat Map')
Out[247]:
In [248]:
# Pipeline for video
def process_image(img):
bboxes = []
# Scale 1
ystart = 400
ystop = 600
scale = 1
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
bboxes = bboxes + rectangles_found
# Scale 1.5
ystart = 400
ystop = 650
scale = 1.5
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
bboxes = bboxes + rectangles_found
# Scale 2
ystart = 500
ystop = 720
scale = 2
rectangles, rectangles_found = find_cars(img, ystart, ystop, scale, color_space, hog_channel, svc, X_scaler, orient, pix_per_cell, cell_per_block)
bboxes = bboxes + rectangles_found
heat = np.zeros_like(img[:,:,0]).astype(np.float)
# Add heat to each box in box list
heat = add_heat(heat, bboxes)
# Apply threshold to help remove false positives
heat = apply_threshold(heat, 3)
# Visualize the heatmap when displaying
heatmap = np.clip(heat, 0, 255)
# Find final boxes from heatmap using label function
labels = label(heatmap)
return draw_labeled_bboxes(img.copy(), labels)
In [249]:
from moviepy.editor import VideoFileClip
from IPython.display import HTML
# Process the test video
video_output = "output_images/test_video.mp4"
clip1 = VideoFileClip("test_video.mp4")
clip1_output = clip1.fl_image(process_image)
%time clip1_output.write_videofile(video_output, audio=False)
In [250]:
HTML("""
<video width="960" height="540" controls>
<source src="{0}">
</video>
""".format(video_output))
Out[250]:
In [251]:
video_output = "output_images/project_video.mp4"
clip1 = VideoFileClip("project_video.mp4")
clip1_output = clip1.fl_image(process_image)
%time clip1_output.write_videofile(video_output, audio=False)
In [252]:
HTML("""
<video width="960" height="540" controls>
<source src="{0}">
</video>
""".format(video_output))
Out[252]:
In [ ]: